[bits-bytes-nn]

Why

Adversarial review rounds R23–R25 covered UX and accessibility (E2E plus edge cases), code and infra hygiene, SEO (Lighthouse and Search Console), and the docs.

Majors found and fixed:

• Mobile LCP was "poor" in lab tests (6.5–14.9 s on the live site). Two causes: render-blocking Google Fonts CSS, and full-size covers sent to phones.
• Wrong citation links in generated posts. Many inline citations pointed at the wrong arXiv paper, and two ids cannot exist. No gate checked external links.
• Fonts: weights nothing uses were requested.
• Docs contradicted the code in places, and scholar-lens's scope was stated wrongly.

Changes

Performance and UX (b81c970)

• Hero covers. <img srcset> with 960/1280/1920 candidates and per-page sizes, fetchpriority=high, decorative (alt="", aria-hidden), hidden in print.
  • script/cover-variants.sh writes the copies. test_every_cover_has_its_phone_and_tablet_copies fails when one is missing or when the original is not 1920px wide.
• Fonts. Google Fonts loads non-blocking (media="print" swap, plus <noscript>), and the serif 400 weights are dropped (font CSS 117 → 93 KB gz).
• CDN failures. GLightbox (JS and CSS) and simple-jekyll-search are async, and the code waits for their load events.
  • A hung jsDelivr previously froze main.js: theme, menu and copy buttons were dead. A hung unpkg left search silent.
  • Now the lightbox is set up only once its CSS has applied, and search gives up on a hung library after 15 s with a message.
• Search index. search.json (about 1 MB gz) is fetched on first use of the search box, with loading and failure status. A query typed earlier still runs.
• TOC at build time. A new _plugins/post_toc.rb filter, with tests, replaces the JS-built TOC. That removes a layout shift after first paint and gives no-JS readers a TOC.
• Accessibility:
  • forced-colors rules for MathJax stretchy brackets (they rendered as solid bars) and for the logo
  • cross-tab theme sync
• SEO: x-default hreflang now points at the English twin, since most search clicks come from English queries.
• CI: weekly external-links.yml runs htmlproofer with external checks; it notifies on failure and blocks nothing.

Lighthouse mobile on local builds, before → after:

• Home: 80 → 100, LCP 4.9 → 1.7 s
• Gemma 3: 76 → 92
• Agentic patterns: 77 → 91

Citation and dead-link fixes (33f73a7). These are hand fixes to generated posts, listed per post. Each old target was checked against the source paper's arXiv HTML bibliography, and every new arXiv id against the arXiv API. An independent reviewer spot-checked 47 ids and about 25 against the bibliographies.

• 2025-02-14-deepseek-coder-when-the-large-language-model-meets-programming-the-rise-of-code-intelligence (1)
  • Li와 연구진: arxiv.org/abs/2207.14255 → arxiv.org/abs/2305.06161 (DeepSeek-Coder text: "Fill-In-Middle (FIM) approach (Li et al., 2023; Bavarian et al., 202)
• 2025-02-27-mistral-7b (1)
  • link: llmboxing.com/leaderboard → web.archive.org/web/20260713010958/https://llmboxing.com/leaderboard (404 with two UAs (site gone). Wayback snapshot 2026-07-13 returns 200 and shows the Mistra)
• 2025-04-04-mixtral-of-experts (3)
  • Kwiatkowski와 연구진: arxiv.org/abs/1901.08634 → aclanthology.org/Q19-1026/ (Mixtral bib [20]: Kwiatkowski et al., "Natural questions: a benchmark for question answeri)
  • Beltagy와 연구진: arxiv.org/pdf/2305.16300 → arxiv.org/pdf/2305.16300 (Mixtral text: "passkey retrieval task introduced in [23]"; bib [23] Mohtashami and Jaggi, )
  • Gao와 연구진: arxiv.org/pdf/2101.00027 → arxiv.org/pdf/2310.10631 (Mixtral text: "a subset of the proof-pile dataset [2]"; bib [2] Azerbayev et al., "Llemma:)
• 2025-10-03-gemma-3-technical-report (9)
  • Chung et al.: arxiv.org/pdf/2305.16264 → arxiv.org/abs/2304.09151 (Gemma 3 bib: Chung et al. (2023) UniMax: Fairer and more effective language sampling for l)
  • Dehghani et al.: arxiv.org/pdf/2309.14322 → arxiv.org/abs/2302.05442 (Gemma 3 bib: Dehghani et al. (2023) Scaling vision transformers to 22 billion parameters, )
  • Chung et al. (2023): arxiv.org/pdf/2305.16264 → arxiv.org/abs/2304.09151 (Gemma 3 bib: Chung et al. (2023) UniMax; arXiv 2304.09151 verified. Old = Muennighoff et a)
  • ZeRO-3: arxiv.org/pdf/1910.02054 → arxiv.org/abs/2101.06840 (Gemma 3 text: 'implementation of ZeRO-3 (Ren et al., 2021)'; bib Ren et al. (2021) ZeRO-Of)
  • JAX: arxiv.org/pdf/2105.04663v2 → arxiv.org/abs/2203.17189 (Gemma 3 text: "'single controller' programming paradigm of Jax (Roberts et al., 2023)"; bi)
  • Mirzadeh et al., 2024: arxiv.org/pdf/2402.09668 → arxiv.org/abs/2410.05229 (Gemma 3 bib: Mirzadeh et al. (2024) GSM-Symbolic, arXiv:2410.05229. Old = How to Train Dat)
  • Shi et al. (2023): arxiv.org/pdf/2305.14031 → arxiv.org/abs/2210.03057 (Gemma 3 bib: Shi et al. (2023) Language models are multilingual chain-of-thought reasoners)
  • Goyal et al. (2022): arxiv.org/pdf/2207.04672 → arxiv.org/abs/2106.03193 (Gemma 3 bib: Goyal et al. (2022) The FLORES-101 evaluation benchmark, ACL; arXiv 2106.0319)
  • Kazemi et al. (2025): arxiv.org/pdf/2411.15124 → arxiv.org/abs/2502.19187 (Gemma 3 bib: Kazemi et al. (2025) Big-bench extra hard, arXiv:2502.19187. Old = Tulu 3.)
• 2025-10-06-tulu-3--pushing-frontiers-in-open-language-model-post-training (7)
  • WildChat: arxiv.org/pdf/2406.18510 → arxiv.org/pdf/2405.01470 (Tulu 3 bib: Zhao et al. (2024), "Wildchat: 1m chatgpt interaction logs in the wild", arXiv)
  • OpenAssistant: arxiv.org/pdf/2310.01377 → arxiv.org/pdf/2304.07327 (Tulu 3 bib: Köpf et al. (2024), "Openassistant conversations-democratizing large language )
  • No Robots: arxiv.org/pdf/2407.12043 → huggingface.co/datasets/HuggingFaceH4/no_robots (Tulu 3 bib: Rajani et al. (2023), "No robots", https://huggingface.co/datasets/HuggingFace)
  • FLAN v2: arxiv.org/pdf/2406.20094 → arxiv.org/pdf/2301.13688 (Tulu 3 text: "FLAN v2 (Longpre et al., 2023)"; bib: "The flan collection: Designing data a)
  • OpenMathInstruct: arxiv.org/pdf/2406.20094 → arxiv.org/pdf/2410.01560 (Tulu 3 text: "OpenMathInstruct (Toshniwal et al., 2024)"; bib: "Openmathinstruct-2: Accele)
  • NuminaMath: arxiv.org/pdf/2406.18495 → huggingface.co/AI-MO/NuminaMath-7B-TIR (Tulu 3 text: "NuminaMath (Beeching et al., 2024)"; bib: "Numinamath 7b tir. https://huggin)
  • Aya: arxiv.org/pdf/2406.20094 → arxiv.org/pdf/2402.06619 (Tulu 3 text: "Aya (Singh et al., 2024b) for multilinguality"; bib: "Aya dataset: An open-a)
• 2025-10-07-deepseek-r1--incentivizing-reasoning-capability-in-llms-via-reinforcement-learning (1)
  • AlphaGo: arxiv.org/pdf/1712.01815 → www.nature.com/articles/nature24270 (R1 text: "Inspired by AlphaGo (Silver et al., 2017b) and AlphaZero (Silver et al., 2017a)")
• 2025-10-07-deepseek-v3-technical-report (2)
  • Byte-level BPE: arxiv.org/pdf/1711.05101 → link removed, text kept (V3 text: "Byte-level BPE (Shibata et al., 1999)"; bib: Shibata et al., "Byte pair encoding)
  • DeepSeek-V2: arxiv.org/pdf/2406.11931 → arxiv.org/pdf/2405.04434 (V3 text: "We adopt a similar approach to DeepSeek-V2 (DeepSeek-AI, 2024c)"; bib 2024c = "D)
• 2025-10-07-gemma-2--improving-open-language-models-at-a-practical-size (3)
  • ZeRO-3 기법: arxiv.org/pdf/2010.04804 → arxiv.org/pdf/2101.06840 (Gemma 2 text: "techniques similar to ZeRO-3 (Ren et al., 2021)"; bib: Ren et al., "Zero-of)
  • Gemma Team의 2024년 연구: arxiv.org/pdf/2305.10403 → arxiv.org/pdf/2403.08295 (Gemma 2 text: "We follow the evaluation setting of (Gemma Team, 2024)"; bib Gemma Team (20)
  • Lin et al.의 연구: arxiv.org/pdf/2024.01234 → arxiv.org/pdf/2401.03315 (Gemma 2 text: "close monitoring of the evolving risks of LLMs (Lin et al., 2024)"; bib: Li)
• 2025-10-11-gemini--a-family-of-highly-capable-multimodal-models (15)
  • Chen et al., 2022: arxiv.org/abs/2204.07118 → arxiv.org/abs/2209.06794 (Gemini bib: Chen et al. (2022) PaLI: A jointly-scaled multilingual language-image model, a)
  • OpenAI, 2023b: arxiv.org/pdf/2303.08774v6 → cdn.openai.com/papers/GPTV_System_Card.pdf (Gemini bib: OpenAI (2023b) GPT-4V(ision) System Card (no URL/arXiv). Old link = GPT-4 Tech)
  • Zhang et al., 2023: arxiv.org/abs/2305.13516 → arxiv.org/abs/2303.01037 (Gemini bib: Zhang et al. (2023) Google USM: Scaling ASR beyond 100 languages, arXiv:2303.0)
  • Chen et al., 2023: arxiv.org/abs/2212.04356 → arxiv.org/abs/2305.18565 (Gemini bib: Chen et al. (2023) PaLI-X: On Scaling up a Multilingual Vision and Language Mo)
  • Barham et al., 2022: arxiv.org/abs/2204.02311 → arxiv.org/abs/2203.12533 (Gemini bib: Barham et al. (2022) Pathways: Asynchronous distributed dataflow for ML (MLSys)
  • Bradbury et al., 2018: arxiv.org/abs/1912.03356 → github.com/jax-ml/jax (Gemini bib: Bradbury et al. (2018) JAX: composable transformations of Python+NumPy program)
  • Dean et al., 2012: www.usenix.org/system/files/conference/osdi12/osdi12-final-74.pdf → proceedings.neurips.cc/paper/2012/hash/6aca97005c68f1206823815f66102863-Abstract.html (Gemini bib: Dean et al. (2012) Large scale distributed deep networks, NeurIPS 25. Old USEN)
  • Leblond et al, 2023: arxiv.org/pdf/2203.07814v1 → storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf (Gemini bib: Leblond et al (2023) AlphaCode 2 Technical Report, URL storage.googleapis.com/)
  • PaLI: arxiv.org/abs/2204.07118 → arxiv.org/abs/2209.06794 (Gemini text: 'PaLI (Chen et al., 2022)'; bib arXiv:2209.06794 PaLI. Old = DeiT III.)
  • Ramesh et al.: arxiv.org/pdf/2021.01000 → arxiv.org/abs/2102.12092 (Gemini bib: Ramesh et al. (2021) Zero-shot text-to-image generation, ICML 2021; arXiv 2102)
  • Zhang et al.: arxiv.org/abs/2305.13516 → arxiv.org/abs/2303.01037 (Gemini text: 'Universal Speech Model (USM) (Zhang et al., 2023)'; bib arXiv:2303.01037 Goo)
  • Google의 데이터센터 네트워크 인프라: arxiv.org/pdf/2203.12533v1 → doi.org/10.1145/3544216.3544265 (Gemini text: 'intra-cluster and inter-cluster network (Poutievski et al., 2022; yao Hong e)
  • JAX: arxiv.org/pdf/1912.03356 → github.com/jax-ml/jax (Gemini bib: Bradbury et al. (2018) JAX, github.com/google/jax. Old 1912.03356 = Cognitive )
  • Universal Speech Model (USM): arxiv.org/abs/2305.13516 → arxiv.org/abs/2303.01037 (Gemini text: 'Universal Speech Model (USM) (Zhang et al., 2023)'; bib arXiv:2303.01037 Goo)
  • Zhao et al.: arxiv.org/pdf/1804.06876 → arxiv.org/abs/2106.08503 (Gemini text: image-to-text people descriptions 'following (Zhao et al., 2021)'; bib Zhao e)
  • GPT-2 report: d4mucfpksywv.cloudfront.net/…/language_models_are_unsupervised_multitask_learners.pdf → cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf (old host no longer resolves; OpenAI's CDN serves the same PDF)
• 2025-10-11-gemini-2.5--pushing-the-frontier-with-advanced-reasoning--multimodality--long-context--and-next-generation-agentic-capabilities (16)
  • Gauthier, 2025: aider.chat/2024/12/21/polyglot.html → aider.chat/docs/leaderboards/ (Gemini 2.5 bib: Gauthier (2025) Aider Polyglot Coding Leaderboard, URL aider.chat/docs/lea)
  • Chowdhury et al., 2024; Jimenez et al., 2024: arxiv.org/pdf/2403.07974 → openai.com/index/introducing-swe-bench-verified/ (Gemini 2.5 bib: Chowdhury et al. (2024) Introducing SWE-bench verified, openai.com/index/i)
  • Rein et al., 2024: arxiv.org/pdf/2403.13793 → arxiv.org/abs/2311.12022 (Gemini 2.5 bib: Rein et al. (2024) GPQA: A graduate-level google-proof Q&A benchmark (COLM)
  • Phan et al., 2025: scale.com/leaderboard/humanitys_last_exam → arxiv.org/abs/2501.14249 (Gemini 2.5 bib: Phan et al. (2025) Humanity's last exam, URL arxiv.org/abs/2501.14249. Old)
  • Wei et al., 2024: github.com/openai/simple-evals → arxiv.org/abs/2411.04368 (Gemini 2.5 bib: Wei et al. (2024) Measuring short-form factuality in LLMs, arXiv:2411.0436)
  • Jacovi et al., 2025: www.kaggle.com/benchmarks/google/facts-grounding → arxiv.org/abs/2501.03200 (Gemini 2.5 bib: Jacovi et al. (2025) The FACTS Grounding leaderboard, arXiv:2501.03200. Ol)
  • Singh et al., 2024: huggingface.co/datasets/CohereLabs/Global-MMLU-Lite → arxiv.org/abs/2412.03304 (Gemini 2.5 bib: Singh et al. (2024) Global MMLU, arXiv:2412.03304. Old = HF dataset page.)
  • Goldman et al., 2025: arxiv.org/pdf/2403.13793 → arxiv.org/abs/2502.21228 (Gemini 2.5 bib: Goldman et al. (2025) ECLeKTic, arXiv:2502.21228. Old = Evaluating Frontie)
  • Balunović et al., 2025: matharena.ai/ → arxiv.org/abs/2505.23281 (Gemini 2.5 bib: Balunović et al. (2025) MathArena: Evaluating LLMs on uncontaminated math )
  • Lee et al., 2024: arxiv.org/pdf/2403.05530 → arxiv.org/abs/2406.13121 (Gemini 2.5 bib: Lee et al. (2024) Can long-context language models subsume retrieval, RAG,)
  • Vodrahalli et al., 2024: arxiv.org/pdf/2403.05530 → arxiv.org/abs/2409.12640 (Gemini 2.5 bib: Vodrahalli et al. (2024) Michelangelo (MRCR), arXiv:2409.12640. Old = Gemi)
  • Yue et al., 2024: arxiv.org/pdf/2403.05530 → arxiv.org/abs/2311.16502 (Gemini 2.5 bib: Yue et al. (2024) MMMU, CVPR 2024; arXiv 2311.16502 verified. Old = Gemini)
  • Padlewski et al., 2024: arxiv.org/pdf/2403.05530 → arxiv.org/abs/2405.02287 (Gemini 2.5 bib: Padlewski et al. (2024) Vibe-Eval, arXiv:2405.02287. Old = Gemini 1.5.)
  • Roberts et al., 2025: arxiv.org/pdf/2403.05530 → arxiv.org/abs/2502.09696 (Gemini 2.5 bib: Roberts et al. (2025) ZeroBench, arXiv:2502.09696. Old = Gemini 1.5.)
  • Conneau et al., 2023: arxiv.org/pdf/2403.05530 → arxiv.org/abs/2205.12446 (Gemini 2.5 bib: Conneau et al. (2023) FLEURS, IEEE SLT; arXiv 2205.12446 verified. Old = G)
  • Wang et al., 2020: arxiv.org/pdf/2403.05530 → arxiv.org/abs/2007.10310 (Gemini 2.5 bib: Wang et al. (2020) CoVoST 2; arXiv 2007.10310 verified. Old = Gemini 1.5.)
• 2025-10-24-qlora--efficient-finetuning-of-quantized-llms (7)
  • T5: arxiv.org/pdf/2210.11416 → arxiv.org/pdf/1910.10683 (QLoRA text: "Super-NaturalInstructions (TKInstruct) [61] with T5 [49]"; bib [49] Raffel et)
  • Flan v2: arxiv.org/pdf/2210.11416 → arxiv.org/pdf/2301.13688 (QLoRA text: "finetuning LLaMA on Flan v2 [39]"; bib [39] Longpre et al., "The flan collect)
  • Alpaca: arxiv.org/pdf/2210.11416 → github.com/tatsu-lab/stanford_alpaca (QLoRA text: "Alpaca [55]"; bib [55] Taori et al., "Stanford alpaca: An instruction-followi)
  • FLAN v2: arxiv.org/pdf/2210.11416 → arxiv.org/pdf/2301.13688 (QLoRA text: "finetuning LLaMA on Flan v2 [39]"; bib [39] Longpre et al., "The flan collect)
  • Vicuna: arxiv.org/pdf/2210.11416 → lmsys.org/blog/2023-03-30-vicuna/ (QLoRA text: "Vicuna [10]" / evaluation protocol of Chiang et al.; bib [10] Chiang et al., )
  • Chiang et al.: arxiv.org/pdf/2210.11416 → lmsys.org/blog/2023-03-30-vicuna/ (QLoRA text: "Vicuna [10]" / evaluation protocol of Chiang et al.; bib [10] Chiang et al., )
  • SwitchBack layers: arxiv.org/pdf/2304.14969 → arxiv.org/pdf/2304.13013 (QLoRA text: "Besides our work, SwitchBack layers [65] is the only work..."; bib [65] Worts)
• 2026-01-12-deepseek-v3.2--pushing-the-frontier-of-open-large-language-models (2)
  • k2-thinking: arxiv.org/abs/2509.07604 → moonshotai.github.io/Kimi-K2/thinking.html (V3.2 text: "(Qwen, 2025; ZhiPu-AI, 2025; MiniMax, 2025; MoonShot, 2025)"; bib MoonShot (20)
  • link: www.anthropic.com/news/claude-4-5-sonnet → www.anthropic.com/news/claude-sonnet-4-5 (Old URL 404 with two UAs; /news/claude-sonnet-4-5 returns 200, title "Introducing Claude S)
• 2026-01-16-constitutional-ai--harmlessness-from-ai-feedback (3)
  • Scheurer et al.: arxiv.org/abs/2212.08073 → arxiv.org/abs/2204.14146 (CAI bib: "Scheurer, J., Campos, J. A., Chan, J. S., Chen, A., Cho, K., and Perez, E. Train)
  • Saunders et al., 2022: arxiv.org/abs/2112.00861 → arxiv.org/abs/2206.05802 (CAI bib: "Saunders, W., Yeh, C., Wu, J., ... Leike, J. (2022). Self-critiquing models for )
  • Gao et al., 2022: arxiv.org/abs/2209.12892 → arxiv.org/abs/2210.10760 (CAI bib: "Gao, L., Schulman, J., and Hilton, J. (2022). Scaling laws for reward model over)
• 2026-04-05-evolution-of-ai-agentic-patterns (1)
  • link: simonwillison.net/2025/Jun/2/the-lethal-trifecta/ → simonwillison.net/2025/Jun/16/the-lethal-trifecta/ (Old URL 404 with two UAs; post "The lethal trifecta for AI agents: private data, untrusted)
• 2026-04-05-evolution-of-ai-agentic-patterns-en (1)
  • link: simonwillison.net/2025/Jun/2/the-lethal-trifecta/ → simonwillison.net/2025/Jun/16/the-lethal-trifecta/ (Old URL 404 with two UAs; post "The lethal trifecta for AI agents: private data, untrusted)
• 2026-04-12-agentcore-harness-engineering-analysis (3)
  • link: simonwillison.net/2025/Jun/2/the-lethal-trifecta/ → simonwillison.net/2025/Jun/16/the-lethal-trifecta/ (Old URL 404 with two UAs; post "The lethal trifecta for AI agents: private data, untrusted)
  • link: www.missioncloud.com/blog/building-enterprise-ai-agents-with-amazon-bedrock-agentcore-lessons-from-a-data-migration-chatbot → web.archive.org/web/20260313041358/https://www.missioncloud.com/blog/building-enterprise-ai-agents-with-amazon-bedrock-agentcore-lessons-from-a-data-migration-chatbot (404 with two UAs; not in missioncloud sitemap (nearest live post "4 AWS Bedrock AgentCore )
  • link: dev.to/kazuya_dev/aws-reinvent-2025-agents-in-the-enterprise-best-practices-with-amazon-bedrock-agentcore-aim3310-2m8a → dev.to/kazuya_dev/aws-reinvent-2025-agents-in-the-enterprise-best-practices-with-amazon-bedrock-agentcoreaim3310-2m8a (Slug typo: dev.to API (articles?username=kazuya_dev) lists the article at ...agentcoreaim3)

Docs (107b0bf)

• README.md and README.ko.md: overview, quick start and a doc map.
• New docs, in English: docs/writing.md, docs/operations.md, docs/architecture.md. They replace docs/tech-doc.md and the long README sections; facts were fact-checked against the code.
• docs/diagrams/category-url-mapping.png now labels the file-name slug, not "title".

_data/post_conventions.yml: comment only (line 8, the docs it points to). No keys changed; scholar-lens's drift check compares parsed data.

Verified

• All five gates, shellcheck and actionlint pass. Unit tests: 187 runs.
• External check: 833 external links, 0 failures, with the new workflow's flags.
• E2E (Playwright, real Chrome):
  • hung and failing CDNs
  • search races (typing before the index or the script)
  • lightbox with late CSS
  • TOC parity with the old JS TOC on all 37 posts
  • forced colors
  • no-JS
• axe-core: 0 violations on home and a post.